2.2 Determining the vocabulary of terms |
27 |
2.2.2Dropping common terms: stop words
STOP WORDS
COLLECTION FREQUENCY
STOP LIST
Sometimes, some extremely common words which would appear to be of little value in helping select documents matching a user need are excluded from the vocabulary entirely. These words are called stop words. The general strategy for determining a stop list is to sort the terms by collection frequency (the total number of times each term appears in the document collection), and then to take the most frequent terms, often hand-filtered for their semantic content relative to the domain of the documents being indexed, as a stop list, the members of which are then discarded during indexing. An example of a stop list is shown in Figure 2.5. Using a stop list significantly reduces the number of postings that a system has to store; we will present some statistics on this in Chapter 5 (see Table 5.1, page 87). And a lot of the time not indexing stop words does little harm: keyword searches with terms like the and by don’t seem very useful. However, this is not true for phrase searches. The phrase query “President of the United States” , which contains two stop words, is more precise than President AND “United States” . The meaning of flights to London is likely to be lost if the word to is stopped out. A search for Vannevar Bush’s article As we may think will be difficult if the first three words are stopped out, and the system searches simply for documents containing the word think. Some special query types are disproportionately affected. Some song titles and well known pieces of verse consist entirely of words that are commonly on stop lists (To be or not to be, Let It Be, I don’t want to be, . . . ).
The general trend in IR systems over time has been from standard use of quite large stop lists (200–300 terms) to very small stop lists (7–12 terms) to no stop list whatsoever. Web search engines generally do not use stop lists. Some of the design of modern IR systems has focused precisely on how we can exploit the statistics of language so as to be able to cope with common words in better ways. We will show in Section 5.3 (page 95) how good compression techniques greatly reduce the cost of storing the postings for common words. Section 6.2.1 (page 117) then discusses how standard term weighting leads to very common words having little impact on document rankings. Finally, Section 7.1.5 (page 140) shows how an IR system with impact-sorted indexes can terminate scanning a postings list early when weights get small, and hence common words do not cause a large additional processing cost for the average query, even though postings lists for stop words are very long. So for most modern IR systems, the additional cost of including stop words is not that big – neither in terms of index size nor in terms of query processing time.
Online edition (c) 2009 Cambridge UP
28 |
2 The term vocabulary and postings lists |
Query term |
Terms in documents that should be matched |
Windows |
Windows |
windows |
Windows, windows, window |
window |
window, windows |
Figure 2.6 An example of how asymmetric expansion of query terms can usefully model users’ expectations.
2.2.3Normalization (equivalence classing of terms)
Having broken up our documents (and also our query) into tokens, the easy case is if tokens in the query just match tokens in the token list of the document. However, there are many cases when two character sequences are not quite the same but you would like a match to occur. For instance, if you search for USA, you might hope to also match documents containing U.S.A.
TOKEN
NORMALIZATION
EQUIVALENCE CLASSES
Token normalization is the process of canonicalizing tokens so that matches occur despite superficial differences in the character sequences of the tokens.4 The most standard way to normalize is to implicitly create equivalence classes, which are normally named after one member of the set. For instance, if the tokens anti-discriminatory and antidiscriminatory are both mapped onto the term antidiscriminatory, in both the document text and queries, then searches for one term will retrieve documents that contain either.
The advantage of just using mapping rules that remove characters like hyphens is that the equivalence classing to be done is implicit, rather than being fully calculated in advance: the terms that happen to become identical as the result of these rules are the equivalence classes. It is only easy to write rules of this sort that remove characters. Since the equivalence classes are implicit, it is not obvious when you might want to add characters. For instance, it would be hard to know to turn antidiscriminatory into anti-discriminatory.
An alternative to creating equivalence classes is to maintain relations between unnormalized tokens. This method can be extended to hand-constructed lists of synonyms such as car and automobile, a topic we discuss further in Chapter 9. These term relationships can be achieved in two ways. The usual way is to index unnormalized tokens and to maintain a query expansion list of multiple vocabulary entries to consider for a certain query term. A query term is then effectively a disjunction of several postings lists. The alternative is to perform the expansion during index construction. When the document contains automobile, we index it under car as well (and, usually, also vice-versa). Use of either of these methods is considerably less efficient than equivalence classing, as there are more postings to store and merge. The first
4. It is also often referred to as term normalization, but we prefer to reserve the name term for the output of the normalization process.
Online edition (c) 2009 Cambridge UP
2.2 Determining the vocabulary of terms |
29 |
method adds a query expansion dictionary and requires more processing at query time, while the second method requires more space for storing postings. Traditionally, expanding the space required for the postings lists was seen as more disadvantageous, but with modern storage costs, the increased flexibility that comes from distinct postings lists is appealing.
These approaches are more flexible than equivalence classes because the expansion lists can overlap while not being identical. This means there can be an asymmetry in expansion. An example of how such an asymmetry can be exploited is shown in Figure 2.6: if the user enters windows, we wish to allow matches with the capitalized Windows operating system, but this is not plausible if the user enters window, even though it is plausible for this query to also match lowercase windows.
The best amount of equivalence classing or query expansion to do is a fairly open question. Doing some definitely seems a good idea. But doing a lot can easily have unexpected consequences of broadening queries in unintended ways. For instance, equivalence-classing U.S.A. and USA to the latter by deleting periods from tokens might at first seem very reasonable, given the prevalent pattern of optional use of periods in acronyms. However, if I put in as my query term C.A.T., I might be rather upset if it matches every appearance of the word cat in documents.5
Below we present some of the forms of normalization that are commonly employed and how they are implemented. In many cases they seem helpful, but they can also do harm. In fact, you can worry about many details of equivalence classing, but it often turns out that providing processing is done consistently to the query and to documents, the fine details may not have much aggregate effect on performance.
Accents and diacritics. Diacritics on characters in English have a fairly marginal status, and we might well want cliché and cliche to match, or naive and naïve. This can be done by normalizing tokens to remove diacritics. In many other languages, diacritics are a regular part of the writing system and distinguish different sounds. Occasionally words are distinguished only by their accents. For instance, in Spanish, peña is ‘a cliff’, while pena is ‘sorrow’. Nevertheless, the important question is usually not prescriptive or linguistic but is a question of how users are likely to write queries for these words. In many cases, users will enter queries for words without diacritics, whether for reasons of speed, laziness, limited software, or habits born of the days when it was hard to use non-ASCII text on many computer systems. In these cases, it might be best to equate all words to a form without diacritics.
5. At the time we wrote this chapter (Aug. 2005), this was actually the case on Google: the top result for the query C.A.T. was a site about cats, the Cat Fanciers Web Site http://www.fanciers.com/.
Online edition (c) 2009 Cambridge UP
30 |
2 The term vocabulary and postings lists |
CASE-FOLDING Capitalization/case-folding. A common strategy is to do case-folding by reducing all letters to lower case. Often this is a good idea: it will allow instances of Automobile at the beginning of a sentence to match with a query of automobile. It will also help on a web search engine when most of your users type in ferrari when they are interested in a Ferrari car. On the other hand, such case folding can equate words that might better be kept apart. Many proper nouns are derived from common nouns and so are distinguished only by case, including companies (General Motors, The Associated Press), government organizations (the Fed vs. fed) and person names (Bush, Black). We already mentioned an example of unintended query expansion with acronyms, which involved not only acronym normalization (C.A.T. → CAT) but also case-folding (CAT → cat).
For English, an alternative to making every token lowercase is to just make some tokens lowercase. The simplest heuristic is to convert to lowercase words at the beginning of a sentence and all words occurring in a title that is all uppercase or in which most or all words are capitalized. These words are usually ordinary words that have been capitalized. Mid-sentence capitalized words are left as capitalized (which is usually correct). This will mostly avoid case-folding in cases where distinctions should be kept apart. The same task can be done more accurately by a machine learning sequence model which uses more features to make the decision of when to case-fold. This is known
TRUECASING as truecasing. However, trying to get capitalization right in this way probably doesn’t help if your users usually use lowercase regardless of the correct case of words. Thus, lowercasing everything often remains the most practical solution.
Other issues in English. Other possible normalizations are quite idiosyncratic and particular to English. For instance, you might wish to equate ne’er and never or the British spelling colour and the American spelling color. Dates, times and similar items come in multiple formats, presenting additional challenges. You might wish to collapse together 3/12/91 and Mar. 12, 1991. However, correct processing here is complicated by the fact that in the U.S., 3/12/91 is Mar. 12, 1991, whereas in Europe it is 3 Dec 1991.
Other languages. English has maintained a dominant position on the WWW; approximately 60% of web pages are in English (Gerrand 2007). But that still leaves 40% of the web, and the non-English portion might be expected to grow over time, since less than one third of Internet users and less than 10% of the world’s population primarily speak English. And there are signs of change: Sifry (2007) reports that only about one third of blog posts are in English.
Other languages again present distinctive issues in equivalence classing.
Online edition (c) 2009 Cambridge UP
2.2 Determining the vocabulary of terms |
31 |
|
Figure 2.7 Japanese makes use of multiple intermingled writing systems and, |
|
;S1 q 6 8 Vc r2 9 T d s3 t: e4 U * u;5 V : & V f! 6 vW<7 " V wgX ) & xF Y h + ! )VG 8 QZ* $ - 9 [H y? ziI% N : {N? ; j* h) : |k: & < +' & l }\( J m = n ) ]> K ; ~:+* ^ ? AL+ @ Mo, +A ?N - B @. ? C ` AO / 4 p0 P - a N ) + D Q 5 bE + R |
like Chinese, does not segment words. The text is mainly Chinese characters with the hiragana syllabary for inflectional endings and function words. The part in latin letters is actually a Japanese expression, but has been taken up as the name of an environmental campaign by 2004 Nobel Peace Prize winner Wangari Maathai. His name is written using the katakana syllabary in the middle of the first line. The first four characters of the final line express a monetary amount that we would want to match with ¥500,000 (500,000 Japanese yen).
The French word for the has distinctive forms based not only on the gender (masculine or feminine) and number of the following noun, but also depending on whether the following word begins with a vowel: le, la, l’, les. We may well wish to equivalence class these various forms of the. German has a convention whereby vowels with an umlaut can be rendered instead as a two vowel digraph. We would want to treat Schütze and Schuetze as equivalent.
Japanese is a well-known difficult writing system, as illustrated in Figure 2.7. Modern Japanese is standardly an intermingling of multiple alphabets, principally Chinese characters, two syllabaries (hiragana and katakana) and western characters (Latin letters, Arabic numerals, and various symbols). While there are strong conventions and standardization through the education system over the choice of writing system, in many cases the same word can be written with multiple writing systems. For example, a word may be written in katakana for emphasis (somewhat like italics). Or a word may sometimes be written in hiragana and sometimes in Chinese characters. Successful retrieval thus requires complex equivalence classing across the writing systems. In particular, an end user might commonly present a query entirely in hiragana, because it is easier to type, just as Western end users commonly use all lowercase.
Document collections being indexed can include documents from many different languages. Or a single document can easily contain text from multiple languages. For instance, a French email might quote clauses from a contract document written in English. Most commonly, the language is detected and language-particular tokenization and normalization rules are applied at a predetermined granularity, such as whole documents or individual paragraphs, but this still will not correctly deal with cases where language changes occur for brief quotations. When document collections contain mul-
Online edition (c) 2009 Cambridge UP